Auction AD System-CTR prediction and Logistic regression

CTR Prediction and Logistic regressionCTR Forecast

From the advertising index we get a lot of ad candidates, these ads logically meet the direction of The Advertiser's condition, that is, the DNF paradigm of the Advertiser orientation condition, in the bidding advertising system, we should choose an optimal scheme, the optimal advertising delivery, the general optimal advertising is through The ECPM is sorted by the generalized second high price, and the charging mode is carried out. In the CPC environment Ecpm=bid * CTR, then the forecast of CTR is the most important part of the auction advertising system.

Each company has a different approach to CTR forecasts, and there are a lot of things to be done here, just to describe everyone's understanding and perception of the problem.

AD Click Predictive probability model:

        The probability of a click can be expressed naturally using conditional probabilities, given A (advertising u (user ), C (page p of Click. We can use statistics to predict the click-through rate, I personally counted as regression than ranking method more appropriate, we know in the search, The method of ranking is extensively researched, and the method of machine learning is used to obtain better sorting results. But we know that the sort of ads is based on ecpm, so the sort of ads can not be considered only ctr, because ecpm also consider bid, in other words, for the advertising network, the bid advertising system needs to accurately estimate ctr. , if it is for DSP, the forecast demand for CTR is higher, because dsp need to simultaneously estimate ctr and click Value, ctr  and click Value determine the dsp bid, that is, to exchange newspaper bid.

The cold-start problem of the new advertisement is also a more difficult problem,Cold-start in the recommendation system research more. The new advertisement in the auction ad system is very many, in the contract type advertisement the new advertisement quantity is not big, because an advertisement's time interval is relatively long, therefore in the contract advertisement can pass the statistic to be possible to solve the new advertisement question better. In the spot ads, every moment there are new advertisers, new ads generated, for these new ads, and there is not much historical data to estimate the effect, but we can use the ad hierarchy (creative, solution, campaign, advertise), and ad tags to estimate new ad clicks. But this approach also has its weaknesses, such as Google has advertisers like Amazon,Amazon to all of its products are advertising, so in this case, the click-through rate of an ad to another ad's CTR does not guide the value.

The auction ad system also captures the dynamic characteristics of clickthrough rates, because a user is willing to create click behavior on an ad, which is fast changing, based on the stage of his shopping, and his instantaneous interest. This is more challenging for display ads than for search ads, where it is entirely necessary to capture the interest of the user, to push the ad, and in search ads, the search term has shown the user's interest and the challenge is much less. From the dynamic characteristics, we want to quickly adjust the characteristics of the feature to capture user interest, on the model, we learn through the online, quickly adjust the model, that is, learning new parameters. Features and models, both of which are dual, for example, if the feature is not changed, then the model will change.

Logistic regression ( Logistic Regression)

The probability of click is subject to bi-nominal distribution, its value can only be 0 or 1, our natural idea is to use the Logistic regression model. Although there are many other more fancy algorithms, it is a simple model that is often used in engineering.

Where x is the n-dimensional feature vector of advertisement,W is the parameter that the last model learns, it can be considered that WTX is a linear function. Generally choose the sigmoid function to squeeze.

I share two perspectives, perspective 1,Logistic regression it belongs to the generalized linear model (generalized linear models), we know that the linear model is the most studied model in mathematics, In fact, the dependent variable is not necessarily linear, and the Logistic regression is a special case of the generalized linear model in binomial error. Perspective 2:Logistic regression is a special case of Maximum entropy model where the number of classes equals 2.

Auction AD System-CTR prediction and Logistic regression